Этот текст также доступен на русском языке: 📈 Как читать графики персентилей
Do you know how to read percentile graphs and actually use them? If not, give me a couple of minutes — you might end up liking them.
A bit of groundwork: how do you read them in the first place? Below are four charts that all show the same metric, just with different scales and different definitions of the “start.”
Let’s imagine we’re looking at how many punnets of strawberries Budapest residents eat in a year. In fact, you can read any data this way; the title is made up. The top‑left chart answers: what percentage of Budapest residents (X) eat no more than Y punnets per year? You’ll see that 27% of the city doesn’t eat any. Consumption then grows exponentially — so much that it’s better to look at the bottom‑left chart: it’s the same chart, but showing only the values for the top 20% of strawberry fans.
The two charts on the right are essentially mirrors of the ones above — the question is flipped: what percentage of Budapest residents (X) eat at least Y punnets per year?
This is why I love percentile graphs when I work with A/B tests: unlike averages, they don’t distort reality. A typical distortion: the mean is meaningless because the top 5% blows it up. On the other hand, the medians do differ — but not by much. A percentile chart shows immediately where the growth happens and where to look closer.
In other words, a percentile chart is useful because it gives you a visual of a hundred statistics at once — each one, in spirit, a cousin of the median.
But my favorite use case isn’t even tests — it’s checking the health of a metric.
Take a look at the gallery:
The curves show roughly the same end values if you only peek at a few key points, but they get there differently. The healthiest curve is slide one: growth between percentiles isn’t perfectly even, but it’s fairly stable. You could easily chalk it up to randomness (and you’d be right: each percentile is just ×1.01–1.012 of the previous one).
Slide two shows a change: neighboring percentiles occasionally stop growing. Slide three goes further: you get chains of five percentiles with no growth.
What does that mean from a product standpoint?
The willingness to buy almost anything often looks like our strawberry punnets:
- Most people buy nothing.
- The top 10–20% generate the bulk of demand.
- In general, each next percentile likely differs from the previous one.
Even strawberry lovers have a constraint: someone doesn’t want 200 or 201 punnets but 200.5. You can’t buy half a punnet, so they buy 200 (because the 201st won’t fit).
What if our product isn’t strawberries but, say, cars? It’s even trickier — you definitely can’t buy half a car. Thankfully, we measure price, so one Lambo owner is “like” ten Teslas or a hundred–two hundred used Moskviches. The catch: you can’t serve demand to the precision of a Moskvich when someone is buying a Lambo. If a buyer wants to spend €500,000, the final trims might be ~€480,000, €490,000, €505,000. That €505,000 no longer fits the €500,000 budget — so the buyer goes for €490,000. The seller just left €10,000 on the table for lack of flexibility.
With cars that’s tolerable — the product is complex and not easily modified. But if you sell something digital and your curve looks like slide three?
Then you’re likely missing money from every customer who ends up buying an option below their true maximum.
Again, why does this happen?
- The customer has a personal maximum they’re willing to pay.
- You offer options above and below that maximum.
- They can’t pay above it, so they pick a lower one.
That’s why a percentile chart is an X‑ray of your revenue. It shows where customers have money — but don’t have a fitting option.
Now you have another tool that can show you more than the usual averages. I hope that next time you look at a jagged chart, you’ll remember the strawberries, the Lambo, and the missing €10,000.